How do AI assistants choose their sources?

Modern AI assistants answer product questions by running a live web search, reading a handful of results, and synthesizing them into one answer with citations. Which sources they pick is decided per question — but in practice they favor pages that directly answer the question: comparison articles, review directories, community threads, documentation, and authoritative brand pages.

Last updated 2026-07-24

Grounding, in one paragraph

Without grounding, a model answers from training data — months old and frozen. With grounding (ChatGPT with search, Claude with web search, Gemini with Google Search, Perplexity always, Grok with live data), the assistant issues real search queries, reads the results, and cites what it used. Grounded answers are what real users increasingly see, which is why measurement without grounding measures the wrong thing.

What tends to get cited

• Comparison and "best of" articles — the single most common citation type for discovery questions.
• Review directories and marketplaces (G2, Capterra, Product Hunt and category equivalents).
• Community discussions — Reddit threads and Stack Exchange answers are heavily cited for "what do people actually use" questions.
• Video — YouTube results appear for how-to and review queries.
• Official product pages and docs — mostly for brand-specific questions, less for discovery.

Why engines disagree with each other

Each engine has its own search index, its own retrieval strategy, and its own synthesis behavior. Perplexity cites aggressively and broadly; Gemini leans on Google’s index; ChatGPT’s citations shift with its search partner data. The same question can put you in three answers out of five — which is exactly why per-engine measurement matters: the gap tells you where to focus.